/* * OPTIMIZED INT16 DDC IMPLEMENTATION * * High-performance polyphase halfband DDC with: * - Symmetric FIR coefficient folding (50% MAC reduction) * - Mirrored ring buffer (zero-overhead wrapping) * - Fused DC+Mix+FIR+Delay processing * * Original work Copyright (C) 2014, Youssef Touil * Modifications Copyright (C) 2025-2026, Benjamin Vernoux * * SPDX-License-Identifier: MIT */ #ifndef IQCONVERTER_INT16_OPT_H #define IQCONVERTER_INT16_OPT_H #include #ifdef __cplusplus extern "C" { #endif /* * Optimized Int16 DDC Context * * Implementation Details: * ---------------------- * - Uses symmetric FIR folding: acc += k[i] * (x[i] + x[N-1-i]) * - FIR buffer is mirrored: write to both [idx] and [idx+BUFFER_SIZE] * - DC removal: 1st-order IIR highpass (alpha = 32100/32768 ~ 0.98) * - Fs/4 mixing: sign pattern [-I, -Q>>1, +I, +Q>>1] * - Q15 arithmetic: coefficients scaled by 32768 * * Memory Layout (64-byte aligned for AVX-512): * - sym_kernel: 16 int32_t (64 bytes) - padded from 12 * - fir_buf: 1024 int32_t (4096 bytes) - 64-byte aligned * - dly_buf: 32 int16_t (64 bytes) - padded from 12 */ #if defined(_MSC_VER) __declspec(align(64)) #endif typedef struct { /* Filter coefficients - padded to 64 bytes (16 x int32_t) */ int32_t sym_kernel[16] #if defined(__GNUC__) || defined(__clang__) __attribute__((aligned(64))) #endif ; /* FIR history buffer (mirrored for wrap-free access) - 64-byte aligned */ int32_t fir_buf[1024] #if defined(__GNUC__) || defined(__clang__) __attribute__((aligned(64))) #endif ; /* Size: 2 * 512 */ /* Delay line for Q-channel - padded to 64 bytes (32 x int16_t) */ int16_t dly_buf[32] #if defined(__GNUC__) || defined(__clang__) __attribute__((aligned(64))) #endif ; /* Circular buffer indices */ int32_t fir_head; /* FIR buffer write position */ int32_t dly_head; /* Delay buffer write position */ /* DC removal state (Q15 fixed-point IIR) */ int16_t old_x; /* Previous input sample x[n-1] */ int16_t old_y; /* Previous output sample y[n-1] */ int32_t old_e; /* Accumulator error term */ /* Delay line length (depends on filter length) */ int dly_len; } #if defined(__GNUC__) || defined(__clang__) __attribute__((aligned(64))) #endif iqconverter_int16_opt_t; /* * Create optimized int16 DDC instance * * Parameters: * hb_kernel - 47-tap halfband filter coefficients (Q15 format) * len - Filter length (must be 47) * * Returns: * Pointer to DDC instance, or NULL on allocation failure * * Notes: * - Extracts symmetric coefficient pairs for folded FIR * - Allocates aligned memory for SIMD optimization * - Initializes all state to zero */ iqconverter_int16_opt_t *iqconverter_int16_opt_create(const int16_t *hb_kernel, int len); /* * Free optimized int16 DDC instance * * Parameters: * cnv - DDC instance (can be NULL) */ void iqconverter_int16_opt_free(iqconverter_int16_opt_t *cnv); /* * Reset optimized int16 DDC state * * Parameters: * cnv - DDC instance * * Resets: * - FIR history buffer to zero * - Delay line to zero * - DC removal state to zero * - Circular buffer indices to initial positions */ void iqconverter_int16_opt_reset(iqconverter_int16_opt_t *cnv); /* * FUSED Processing: uint16_t ADC samples -> int16 I/Q output * * Uses precomputed LUT for ultra-fast ADC -> int16 conversion. * * Processing Pipeline (Fused): * --------------------------- * For each 4-sample ADC block [x0, x1, x2, x3]: * * 1. LUT-based ADC Conversion + DC Removal (1st-order IIR highpass): * x_int16 = lut[adc_sample] * y[n] = x[n] - x[n-1] + 0.98*y[n-1] * * 2. Fs/4 Mixing with center tap scaling: * I: [-x0, +x2] -> FIR path (I-channel) * Q: [-x1>>1, +x3>>1] -> Delay path (Q-channel) * * 3. Symmetric FIR (I-channel): * acc = sum k[i] * (x[i] + x[23-i]) for i=0..11 * * 4. Delay compensation (Q-channel): * Delay = (47-1)/4 = 11.5 -> 12 samples * * 5. Output (complex I/Q): * [I_filtered0, Q_delayed0, I_filtered1, Q_delayed1] * * Performance: * ----------- * - ~400 MSPS on modern x86_64 (vs ~170 MSPS legacy) * - ~2.4x performance improvement over legacy */ void iqconverter_int16_opt_process_u16(iqconverter_int16_opt_t *cnv, const uint16_t *src, int16_t *dest, int len); /* * FUSED Processing: uint8_t ADC samples (8-bit mode @ 40 MSPS) -> int16 I/Q output * * Same processing pipeline as _u16 variant but with 8-bit ADC input. * Uses 8-bit LUT (256 entries) for ADC -> int16 conversion. * * Parameters: * cnv - DDC instance * src - Raw uint8_t ADC samples (input) * dest - int16_t I/Q output buffer * len - Number of input samples (must be multiple of 4) */ void iqconverter_int16_opt_process_u8(iqconverter_int16_opt_t *cnv, const uint8_t *src, int16_t *dest, int len); #ifdef __cplusplus } #endif #endif /* IQCONVERTER_INT16_OPT_H */